Computational methods for the detection of somatic structural variants in cancer genomes using long-read sequencing

Abstract

Accurate detection of somatic structural variants (SVs) is critical for informing the diagnosis and treatment of human cancers. In this thesis, I present SAVANA, a computational method for the analysis of somatic SVs using long-read whole genome sequencing data from tumours and matched normal samples. SAVANA employs machine learning to distinguish true somatic SVs from germline events and noise. Additionally, I establish best practices for benchmarking SV detection using simulated and sequencing replicates to demonstrate SAVANA’s superior sensitivity, specificity, and speed compared to existing methods. I show that SAVANA performs robustly across a variety of clonality levels, genomic regions, SV types, and sizes. Using Illumina and Oxford Nanopore whole-genome sequencing data from 99 tumours and matched normal samples of patients, I show that SVs reported by SAVANA are highly concordant with those detected using short-read sequencing, including in regions of complex structural variation. I also highlight the enhanced ability of long-reads to identify SVs in repetitive regions where short-reads are unable to map with high confidence. In summary, this thesis introduces SAVANA as a novel computational method to identify somatic SVs in long-reads, establishes a robust framework for benchmarking SV detection, and demonstrates the method’s high consistency and enhanced sensitivity compared to short-reads across a large patient cohort.